Graph embedding learning method based on graph primitives

Abstract

The invention provides a graph embedding learning method based on graph primitives, and the method comprises the steps: enabling the graph primitives (Motif) of different structures to serve as supernodes to be inserted into a network, and constructing a heterogeneous network which consists of Motif super nodes and original nodes in a graph; in order to generate a corpus from a heterogeneous network, providing a random walk strategy based on Motif to ensure that nodes with high connectivity or high structural similarity are relatively close to each other in the corpus. According to the method, not only can the embedded vectors of the nodes in the graph be learned, but also the embedded vectors of different graph primitives Motif can be obtained; by means of the embedded vector of the graph primitive Motif obtained through learning, the inductive learning problem under the dynamic scene can be effectively solved; for newly-entered nodes in the dynamic graph, the learned graph primitiveembedding vector algorithm is used, the frequencies of various primitives can be rapidly counted around the new nodes, and the embedding vectors of the newly-entered nodes are rapidly calculated through weighted summation.

Description

technical field [0001] The invention relates to the technical field of graphics processing, in particular to a graph-based element-based graph embedding learning method. Background technique [0002] In daily life, graph-structured data is widely used. In the field of biochemistry, molecules can be regarded as a graph, and each atom is regarded as a node, connected by chemical bonds; in the academic citation network, published papers or scholars are connected to each other through references to each other; in the field of e-commerce, recommendation systems can Efficient and accurate recommendations are made based on graphs based on user and product composition. The graph structure is irregular: nodes may have different numbers of neighbors, nodes themselves may have complex characteristics, and edges may have various forms, such as directed edges, undirected edges, authorized edges, unauthorized edges, and so on. Compared with other structures, the graph structure can conv...

AI Technical Summary

Problems solved by technology

[0004] However, real-world data are very complex and noisy

The above known models have the following deficiencies: 1) Using a single similarity metric cannot capture the information features in the graph well, which will lead to a certain loss of information. For example, the algorithm based on structural similarity is not suitable for graphs with strong homogeneity. data, e.g. users with the same interests are more likely to connect than randomly, and algorithms that only consider connectivity will fail in tasks that value structure

2) In the past, these models were only suitable for static graphs, and could not handle inductive learning tasks in dynamic scenarios

3) The past model cannot further generate a representation vector for a certain group while learning the graph embedding vector. For example, in a social network, the graph element motif of a closed triangle indicates that node users often introduce friends to each other, while the existing model Unable to explicitly and quantitatively learn to represent this behavior in social situations or the group that tends to introduce friends

Images